How word solvers work, in plain English

Every tool on this site runs entirely in your browser, using the same dictionary and a few simple techniques. Nothing here is a black box, and nothing you type is sent to a server. This guide explains exactly what happens between pressing the button and seeing results, with the real numbers from our word list.

It starts with a good word list

A solver is only as good as its dictionary. Ours is built from ENABLE, a public-domain word list created for word game software, plus a short supplement of words whose acceptance in current North American play is well documented (QI, ZA, FE, KI, OI, OK and EW). The combined list has 172,830 words. Proper nouns, abbreviations and hyphenated terms are excluded, just as they are in the games themselves. Our editorial standards page explains the trade-offs.

Word lengths are not evenly spread. Short words are few and long words are rare, with a peak in the middle:

The longest entry is ETHYLENEDIAMINETETRAACETATES, at 28 letters. The dictionary is downloaded once, the first time you use a tool, and then kept by your browser, which is why later searches feel instant.

Why trying every arrangement fails

The obvious way to unscramble letters is to try every ordering and look each one up. That falls apart quickly. Seven letters can be arranged 5,040 ways, ten letters 3,628,800 ways, and twelve letters about 479 million ways. Worse, you also want shorter words made from some of the letters, which multiplies the work again. No phone wants to do that.

Counting letters instead

Solvers flip the problem around. Instead of generating arrangements, they walk through the dictionary once and ask a much cheaper question of each word: could this word be built from these tiles?

To answer it, the solver counts letters. For the rack RETAINS, the count is one each of R, E, T, A, I, N and S. For each dictionary word, it counts that word's letters and checks that no letter is needed more times than you have it. In outline:

count the letters in your rack
for each dictionary word no longer than the rack:
    for each distinct letter in the word:
        if the word needs more copies than the rack has:
            reject the word
    otherwise keep it, and work out its score

That is a few hundred thousand cheap comparisons, which a browser finishes in roughly 50 milliseconds in our own tests. RETAINS, for example, produces 256 words, nine of which use all seven letters.

Blank tiles are a budget

A blank can stand in for any letter. The solver treats blanks as a budget: when a word needs a letter you do not have, it spends one blank to cover it. If the word needs more substitutions than you have blanks, it is rejected.

Scoring has to respect the same rule, because a blank is worth zero points. Take the rack QUIETS? and the word BISQUE. Counted naively, BISQUE is worth 17 points. But you have no B, so the blank becomes the B, and the real score is 14. Our tools always show the real figure. Getting this wrong is one of the most common bugs in word finders, and it makes the "best" play look better than it is.

Anagrams and letter signatures

An anagram solver uses the same letter counting with one extra rule: the word must be exactly as long as your rack. That single check reduces the candidates dramatically.

There is a neat trick behind finding anagram families quickly. Sort a word's letters alphabetically and you get its signature: LISTEN, SILENT and TINSEL all become EILNST. Two words are anagrams exactly when their signatures match. Group the dictionary by signature and every anagram family appears at once. That is how we found the bingo stems in our Scrabble strategy guide: we tested which seventh letters turn each six-letter signature into a real seven-letter word.

Patterns for crosswords

For crosswords, the solver converts your pattern into a regular expression, a compact description of what a match must look like. Each underscore or question mark becomes "any one letter" and each star becomes "any run of letters". The pattern c_t_g_r_ becomes "c, any letter, t, any letter, g, any letter, r, any letter". When the pattern has no star, the solver only looks at words of the right length, which is a big shortcut. Then the optional filters (must contain, cannot contain, starts with, ends with) remove anything that does not fit.

Wordle is a set of constraints

The Wordle Solver turns each colored tile into a rule. Green fixes a letter in a position. Yellow says the letter is present but not in that position. Gray caps how many times the letter may appear: zero, unless the same letter is also green or yellow elsewhere in the guess, in which case the cap is that count.

Every five-letter word in the dictionary is checked against every rule, and the survivors are ranked. The ranking favors words whose letters appear in many of the other remaining candidates, because guessing such a word tends to split the remaining field well. In our Wordle strategy guide we take the idea further and measure how many words each starting guess leaves on average.

What solvers cannot do

• Multi-word anagrams. Splitting letters into two or more words multiplies the possibilities enormously, and most results are nonsense phrases. We stick to single words.
• Reading clues. The crossword tool matches patterns. It does not understand clues, which is the fun part you get to keep.
• Knowing your board. The Scrabble and Words With Friends tools score tiles, not positions. Premium squares depend on where you play.
• Official lists. Tournament dictionaries are licensed and differ slightly from ours, so unusual words should be checked against your event's list.

How we test the tools

Word tools look simple, which is exactly why small mistakes slip through. We check ours in three ways.

Known answers. We keep a set of racks and patterns whose correct results we have worked out independently, such as the six anagrams of LISTEN (ELINTS, ENLIST, INLETS, LISTEN, SILENT and TINSEL) and the single eight-letter match for c_t_g_r_. After any change, the tools must reproduce them exactly.

Edge cases. The bugs worth worrying about live at the edges: blank tiles, repeated letters, and racks with no answers at all. Two real examples from our own testing: an early version showed each word's full tile value even when a blank supplied one of its letters, so BISQUE appeared to beat words it actually trails. Another version of the Wordle board quietly dropped letters typed quickly. Both were caught by deliberately feeding the tools awkward input, and both are now part of the checks.

Claims in the text. Any word our guides cite as valid is checked against the dictionary automatically before a page goes live, and statistics such as word counts, letter frequencies and bingo stems are calculated from the dictionary at build time rather than typed in by hand. If the word list changes, the numbers on the page change with it.

If you find a result that looks wrong, please tell us. Every fix is listed on our What's new page.

Privacy by design

Because all of this runs on your device, your letters, patterns and guesses never leave your browser. There is nothing for us to log or store. The only network request the tools make is the one-time download of the word list itself.

Questions readers ask